MSEDoA

📍 This is an official PyTorch implementation MSEDOA: Enhancing DOA Estimation with Multiscale Squeeze-and-Excitation Networks for Automotive Millimeter-Wave Radar

📰 News

[2024.1126] Paper accepted and published.

[2024.0830] Model checkpoints and training logs are released! best_model.pth should be placed incheckpoints/Model20240825181151 to reproduce the experimental results.

[2024.0825] Code of MSEDOA are released!

💡 Methodology

For DOA estimation, the MSEDOA framework integrates multi-scale feature extraction, channel attention mechanisms, and SE-ResNet, all of which are essential for capturing the complex nonlinear relationships between raw signals and their corresponding DOAs.

🛠️ Setup

Step1: create a new conda environment.

git clone https://github.com/Armorhtk/MSEDOA.git
cd MSEDOA
conda create -n MSEDOA python==3.9.0
conda activate MSEDOA
pip install requestments.txt

Tips: We test the framework using pytorch=2.4.0, and the CUDA compile version=12.2 Other versions should be also fine but not totally ensured.

Step2: Generate simulation data.

If you need to use the same dataset as in the paper, you can download the 1.27GB data_ongrid_250k dataset from the Baidu Cloud Link.

python data_gen.py 
  --folder_path './data_ongrid_100k' \
  --number_elements 12 \
  --val_num_samples 180000 \ 
  --val_num_samples 20000 \
  --max_targets 3 \
  --snapshot 10 \
  --min_angle -60 \
  --max_angle 60 \
  --SumlationSEED 42

The data is stored in the data_ongrid_100k folder with two sub-folders train and Val for the signal and label tensor data under different SNR.

⏳ Training

python train.py 

Run train. py to obtain the best model. Notice: To reproduce completely consistent results, load the released weights of the best model in checkpoints folder and proceed directly to the Evaluation stage

🔖 Evaluation

Experiment A: Testing the DOA estimation performance of the DOAMOE method. The finished program allows you to view the results in the image folder. Set the signal-to-noise ratio (SNR) to 12 dB. $\theta_1$ ranges from [-60, 55] degrees, and $\theta_2$ is 4.7° apart from $\theta_1$. Test the performance and error of different DOA estimation methods on the same simulation data.

python Evaluation_Performance.py 

DOA Performance	DOA Errors

Experiment B: Testing the DOA estimation performance of difference SNR. The finished program allows you to view the results in the image folder. Set the signal-to-noise ratio range -20 to 30 dB. $\theta_1$ ranges $[-60^{\circ},53^{\circ}]$ to create a more realistic continuous angular distribution, and $\theta_2$ is 6.1° apart from $\theta_1$. Test the performance and error of different DOA estimation methods on the same simulation data.

python Evaluation_SNR4RMSE.py

📜Citation

If you find this work helpful for your project,please consider citing the following repository:

@inproceedings{hu2024msedoa,
  title={MSEDOA: Enhancing DOA Estimation with Multiscale Squeeze-and-Excitation Networks for Automotive Millimeter-Wave Radar},
  author={Hu, Tingkai and Sun, Shuang and Wu, Zhenyu and Li, Chuandong and Zhan, Choujun and Xiong, Hailing and Luo, Zhen},
  booktitle={2024 IEEE International Symposium on Product Compliance Engineering-Asia (ISPCE-ASIA)},
  pages={1--4},
  year={2024},
  organization={IEEE}
}

🙏 Acknowledgement

• SENet